LAPUSKA: Fast Image Super-Resolution via LAPlacian UpScale Knowledge Alignment

Abstract

Single image super-resolution is important part of computer vision open problems. Recently, deep neural networks have demonstrated excellent performance in this problem. In this work, several cutting edge methods for super-resolution problem using deep neural networks will be considered. Comparison of their effectiveness and evaluation of neural networks architectures with respect to different metrics is one of our main goals for this research. Modern deep learning methods often require large computational cost and load a lot of computer memory, which affects the ease of use of neural networks and the time of generation super-resolution results. In addition to the existing models, we propose a new architecture of neural networks based on best properties of considered architectures and designed to eliminate their shortcomings. Furthermore, we compare the quality of all considered deep learning methods with baseline method of bicubic interpolation.